0:00 No, it was it was really great to hear the last talk. I was I was at Carbon before we started C-Infinity and at Carbon I was working on design engine which Andrew talked about which is very cool to see how how far they’ve gone. So at C-Infinity we’re we’re thinking u you know there’s already so much challenges so many challenges in making parts as you saw from the last couple of conversations and designing a part can involve a lot of analysis of physics but and you know customization but a lot of challenges still remain when you go from parts to assemblies you know and so what we’re looking at at C-Infinity is really thinking about how to how to overcome the complexity of manufacture mfacturing assemblies.

0:42 And you know, one of the things that people talk about a lot right now is physical AI. And physical AI is basically autonomous hardware operating in the real world. And so one of the, you know, biggest use cases of physical AI in manufacturing is in robotic assembly and how robots can actually take complex products, customized products and assemble them together. But the challenge is that the prominence of these assemblies comes from design and designs are created in 3D CAD systems as you all know and you have a digital 3D assembly and the output of the design process for for an assembled product is what is called an engineering bill of materials or an Ebomb right like this is the standard terminology they use in product life cycle management.

6:51 So, how does the Ebon actually go into the set of instructions that actually are used on the factory floor to be able to make the assembly? And that’s a really messy process today. And it’s a it’s a problem that we solve called process planning. And fundamentally, it’s a problem about how to convert one structure, one representation of a product into another representation. So, the Ebon is the engineer’s view of the product.

It’s a functional breakdown of the product into sub assemblies that typically have their own distinct functions like a hydraulic system or pneumatic system or structural system. But that needs to be converted into the manufacturing bill of materials called the Mbomb. And today that process is entirely manual, right? You use PLM systems and you use a lot of you know very very archaic techniques like creating exploded views and powerpoints and screenshots and it’s really really painful and it’s messy.

So solving this problem of converting the E-bomb to the M bomb is quite fundamental in manufacturing but it is also something that we abstracted about and saw that it is a sort of a general general problem that we must solve. Whenever when there is intent that needs to be converted to instructions that’s what a compiler does. And if you look at software programs, you you represent your intent in terms of a programming language.

And a compiler works by taking that that instru that representation of your intent, mapping it to your architecture on which you’re actually going to execute your instructions and creates that sequence. So what we’re building at C-Infinity is a protocol auto assembler which we’re calling it as a compiler for physical artificial intelligence. Okay. So I’m going to talk about how we do that now. So after we started C-Infinity, we spent a whole bunch of time with manufacturers at scale like so we work with some of the Fortune 100 manufacturers and people who are are really advanced and and savvy in terms of how to bring modern techniques into manufacturing, modern software techniques into manufacturing.

But you still see that like there is so much that needs to be done. Designs that come out of you know various design departments that need to go into manufacturing they’re filled with bugs right like there are all kinds of unintended interferences there’s alignment issues there’s all kinds of you know issues going on. So if you think of the analogy to a program, right? Like that program is basically filled with all kinds of errors, right?

Like and you need one of the things a compiler can do for you is to be able to uncover those errors and to understand like where you have let’s say syntax errors for example, right? Like in your design and so debugging source CAD early and quickly is something that is very valuable to a design engineer. And and it avoids a lot of painful iterations. The process today is you go through a design review cycle and you uncover these issues sometimes but you want to uncover them early.

The other challenge is that like in CAD systems and in the design process typically we start from parts and we go all the way up to assemblies you know and so it’s a bottomup process and you know the assembled products typically are what you get out as the final product and so discovering issues in the design at the assembly stage can often force really expensive redesigns and and triggers all kinds of emergency you know design reviews and so on.

So you want to be able to overcome that and identify issues early. The other the other problem we’ve seen is that products you know are increasingly customized and configured like we heard from the previous conversations. And so this is happening at the assembly level too, right? Like the simplest example is if you want to buy a computer, you know, you can configure your GPU, you can configure your, you know, your hard drive and whatever, right?

9:24 Like so you can do that not just for for simple consumer products, but you can also do it for like really complex like off-road automotive products and semiconductors and so on. So when you have multiple configurations of a product, every change you make to a design ripples across all of those configurations, you know, and and they can affect the manufacturability of a whole product family. Now you’re you’re thinking about not just a single design, you’re thinking about a class of designs.

So how do you manage that kind of complexity? So we said that like if you can actually constructively define a process plan from a design take take a design as the input and generate a sequence of operations that can be used to make that design or assemble that design. In our case it is a constructive proof of design for assembly. And so if you can construct that proof then you know you at least have a valid assembly sequence or an a valid process plan.

But if you can’t, you have an issue that now going from design to manufacturing, you want to be aware of how the products come together in your specific factory. Like if a company has, you know, production lines in say Indiana versus Brazil or Germany, right? Like those factories are going to look different and the way material flows in each of those factories is also going to be different.

So you want to be able to capture that information as a part of your your compilation target as I call it you know so how does auto assembler work right like so it works inside your your CAD program and here I’m showing Creo for example and so what happens is you take an you take your 3D model and you upload it into auto assembler inside your CAD program and it comes up with an automated sequence of operations right of installation operations and so it’ll it’ll basically basically come up with the the it it figures out the relationships between the part geometries, understands how they can be moved relative to each other, understands the precedence relationship of what needs to come before what and organizes all of that into a into a coherent process plan.

So to show how that works out right like so if you if you create a new plan and in order assembler you can create multiple what if scenarios. This is something that design and manufacturing engineers love to have right like so they can collaborate and have a conversation about how to converge to the right plan. So in a in a whatever scenario you can in a new scenario you can you know construct a plan for each sub assembly and you can you can customize your plan to be able to create specific groupings of parts.

Groupings is a way in which we we do design like for example you have bolt patterns and and mirror patterns that you see all the time in design geometries but you also use it in manufacturing to be able to convey the right kind of instructions you know so similar parts need to be treated similarly you know in a process plan. So we can create that level of of you know specificity inside the inside the process plan.

The planner is fast and it adapts immediately to be able to take in human input and recomputee a process plan that actually works. And the next challenge in the ebomb momb conversion problem is you know you have to be able to adapt your your structure to the manufacturing line. So in this case oops what I’m saying is that u let me just go back here. What what we’re trying to show here is that if you look at the the structure of the assembly, I want to make sure in in the designer has said that the timing cover and the water pump, these two subasssemblies over here are are distinct subasssemblies, but the manufacturer says, you know what, I’m going to take all of these parts together as a bag of parts and make them together in a single assembly cell.

Right? So, how does that work out in auto assembler? So you can you can move your your design structure and reorganize it and try to construct a manufacturing plan, right? Like so you do that and you you construct a plan for the water pump and the timing cover, you get a new plan. Now here’s an example of a bug, right? Like that happens. So where what you see is that like you have a bunch of parts that belong to a bearing that are shown, you know, in red.

9:43 So why is that? So auto assembler says that well you know what like the the design of this bearing has the balls of the bearing inside the inside the bearing race and there’s no way in which I can assemble this. So but what happens in this case is that this is actually a part that I got from the supplier right and the suppliers they don’t give you entire CAD models.

10:03 So you can say you know this is a preassembled geometry don’t try to construct a plan for this and auto assembler is going to you know work with those constraints and and create a plan for the rest of it and the third thing we see from from engineers or requests from engineers that they want to customize plans. So they say, “Yeah, this is great, but like I want to take, you know, the the the shaft over here and I want to move it all the way up right after the impeller because I want to use the shaft to align the parts that come after that in the assembly sequence.” So you can you can reorder the plan if you like you know so it’s like a human AI partnership over here where the human engineer says like okay the these these steps in the plan need to be updated and the artificial intelligence solver can figure out whether that plan is valid or not and comes up with the u u with with the with the right sequence and the the the next thing that automate does is it’s introspective right like so it looks at it looks at what it can do with the plan and it says you know there are some places where I just simply can’t construct a plan and I see some interferences here but maybe you know how to solve it human right like because it might involve some complex motion of some really flexible geometry that I can’t really figure out in a in a meaningful way so the human can override the AI’s roadblock effectively and then the AI progresses with the plan to come up with the updated sequence so there’s There’s all kinds of things that we do inside auto assembler to build this collaborative intelligence between the the human engineer and the automatic planner that that become very very useful to the process of converting the Ebomb to the Mbomb.

11:52 So we call this building AI with mechanical intuition. The output of auto assembler is a set of instructions that are directed either towards humans or or to robots. Human output, human output is presented in the form of work instructions and robotic output is presented in the form of joint angles and programs that you can use to actually construct the sequence. So this is it’s mechanical intuition because we take the design and infer the design intent from it.

12:22 We figure out how the parts were supposed to move together or interact with each other from just a geometry specification. Converting the language of geometry to the language of motion is a is a challenging problem to solve and we were able to overcome some of those challenges and you see that it is a collaborative tool you know so you can you can work together with the AI to construct your process plans and output the the instructions and this it’s been a year since we started a year and a half since we started the company and the solution is already deployed in some leading manufacturing companies including a couple of Fortune 100 companies, right?

13:02 Like so they’re using it as a part of their their day-to-day process. So so the challenge in this Ebomb to Mbomb conversion can be seen in this diagram, right? So the this is the engineering design representation of the assembly tree and this is the manufacturing representation, right? Like of how the the parts need to come together in sequence and the the the transformations is what we can do with auto assembler and now we get to the next level right like configurations product configurations and how do we manage them so as I said before there’s a lot of customization for for hardware products and in the in the picture above I have three examples of a fairly simple assembly that has three distinct this existing configuration.

13:52 So variants of products are created all the time for customization. But from a manufacturing point of view, you want to reuse existing knowledge as much as you can, right? Like so you if you have a product where you know some of the subcomponents or sub assemblies are are reused, you don’t want to redesign or recomputee a manufacturing plan or an Mbomb for those for those subcomponents and you want to be able to take advantage of what you’ve done before.

14:18 And this is especially true in high mix manufacturing. Right? So this is where we get into the concept of a configuration space. Right? So so in a lot of early in the early days of robot motion planning. What people tried to do was to define what is called a configuration space which is basically all the possible states in which a robot can can move and and try to partition it into into regions where the robot cannot move versus where the robot can move.

14:48 Move and then figure out a motion plan you know in that space. So this is the same concept that can be applied to to product assemblies, right? So if you have varants of products, you want to think about like how they can be organized in such a way that some design modifications are in feasible and whenever the modifications are feasible, you want to reuse existing knowledge. And what we we do to solve that problem is to use machine learning to create a representation of your entire product variant space.

15:21 So one of the things you can do very well with machine learning is create embeddings of geometry. And when you have embeddings of geometry represented as a data set then you can then as a database then you can query and pull from existing geometries to do surgery as I call it. So you have like you you pick and place the the components of a plan that you need to and stitch it together to create your your new mbomb representation.

So this allows users to generate output 10x faster than they can today. Right? So this is a really effective procedure. I’ll go really quickly through this video to show you what the u the reuse looks like. So here’s one variant of a design, you know, and we’ve constructed a plan for it. And once you you’ve constructed this plan and you save it, the next variant is going to is going to look at it’s something is is something that’s very similar to what you’ve seen before except the piston, you know, you know, in a different configuration and it reuses the plan for the piston and the back plate in this case.

16:31 And it only it only plans for the things that you actually have not seen before. So reuse and reconfiguration are two big parts of how you can speed up manufacturing planning. And ultimately how underass assembler works is that like it takes in a design, takes in a library of existing plans which reflects the topology of the layout of your shop floor. Takes that together, constructs a process plan after giving you design for assembly feedback and the compiler’s output which is an M1 is then fed back into the library to learn continuously about your enterprises you know knowhow essentially and then you you compile it further down into factory specific instructions.

So the the the reason why this is powerful is because you can deploy and we deploy in virtual private clouds. We don’t cross trainin on manufacturers data. So you on once you you run it a few times and you see mbomb output the the quality of the plans just starts improving rapidly after that and I want to connect this back upstream why we started solving this problem.

17:41 Not only is it a pain point today and where where this is going is upstream towards design right and yes planning is a is a constructive proof of design for manufacturability yes it’s a pain point where I see the the real advantage of this approach is in in simplifying complex systems and one of the ways to simplify complex systems is through parts consolidation and when you try to figure out how to reduce cost or reduce the number of parts may not always be the same But if you’re trying to reduce the number of parts and and create these nonlinear system improvements, how do you identify the collection of parts that needs to be consolidated?

18:21 Planning gives you insight into that. If a collection of parts can be put together in the same assembly sequence, right? And if they have similar material properties, then that’s a good candidate for you to think about in terms of parts consolidation. And and thinking about it one step further, maybe you uncover new manufacturing processes like how Tesla did with the Giga Press, right? Like they took the undercarriage of the Model Y and from 300 cars they made it I guess two parts and now one part using a stamping machine, right?

18:50 Like a large scale stamping machine. So, so this this coupling between design and manufacturing can only be identified if you look at the assembly level, if you look top down into the manufacturing process. And so that’s kind of where we’re seeing this go. So we want to un unlock system level efficiency, search a product configuration space, compile intelligently to get process plans to transform the Ebon to the Mbomb.

19:16 This is a it’s a huge problem $50 million year over year for large manufacturers, you know, and and get to the point where we can compile it all the way to physical AI and go upstream to design. That’s our vision for C-Infinity. If you’re interested in talking more about auto infinity, please talk to me after this. This is a soon to be launch, but it’s not quite ready. That’s good because next time we have dummy on talk about lab.